Document 10938501

advertisement
Hindawi Publishing Corporation
Journal of Inequalities and Applications
Volume 2007, Article ID 90526, 10 pages
doi:10.1155/2007/90526
Research Article
Schur-Type Inequalities for Complex Polynomials with
no Zeros in the Unit Disk
Szilárd Gy. Révész
Received 20 March 2007; Accepted 28 June 2007
Recommended by Saburou Saitoh
Starting out from a question posed by T. Erdélyi and J. Szabados, we consider Schurtype inequalities for the classes of complex algebraic polynomials having no zeros within
the unit disk D. The class of polynomials with no zeros in D—also known as Bernstein or Lorentz class—was studied in detail earlier. For real polynomials utilizing the
Bernstein-Lorentz representation as convex combinations of fundamental polynomials
(1 − x)k (1 + x)n−k , G. Lorentz, T. Erdélyi, and J. Szabados proved a number of improved
versions of Schur- (and also Bernstein- and Markov-) type inequalities. Here we investigate the similar questions for complex polynomials. For complex polynomials, the above
convex representation is not available. Even worse, the set of complex polynomials, having no zeros in the unit disk, does not form a convex set. Therefore, a possible proof
must go along different lines. In fact, such a direct argument was asked for by Erdélyi and
Szabados already for the real case. The sharp forms of the Bernstein- and Markov-type
inequalities are known, and the right factors are worse for complex coefficients than for
real ones. However, here it turns out that Schur-type inequalities hold unchanged even
for complex polynomials and for all monotonic, continuous weight functions. As a consequence, it becomes possible to deduce the corresponding Markov inequality from the
known Bernstein inequality and the new Schur-type inequality with logarithmic weight.
Copyright © 2007 Szilárd Gy. Révész. This is an open access article distributed under the
Creative Commons Attribution License, which permits unrestricted use, distribution,
and reproduction in any medium, provided the original work is properly cited.
1. Introduction
Let ᏼn and ᏼcn denote the set of one-variable algebraic polynomials of degree at most n
with real, respectively, complex coefficients, and denote the set of all the (real or complex) polynomials by ᏼ and ᏼc , respectively. The open unit interval will be denoted by
2
Journal of Inequalities and Applications
I := (−1,1), and the open unit disk {z : |z| < 1} will be denoted by D. We take
f := sup | f |
(1.1)
I
as the norm of a polynomial or a continuous function.
In approximation theory, Schur- and Bernstein-type polynomial inequalities constitute an important subject, see, for example, [1, 2]. The classical inequality of Schur states
that
p ≤ (n + 1) p(x) 1 − x2 p ∈ ᏼn .
(1.2)
α
This can be generalized to weights (1 − x2 ) with α > 0 as well:
α p ≤ C(α)n2α p(x) 1 − x2 p ∈ ᏼn .
(1.3)
Schur’s inequality (1.2) is usually combined with Bernstein’s inequality
p(x) ≤ √ n
p
1 − x2
p ∈ ᏼn
(1.4)
to deduce Markov’s inequality
p ≤ n2 p
p ∈ ᏼn .
(1.5)
Not only Markov’s inequality but also many other results hinge upon the basic inequalities of Schur and Bernstein. Thus, there is a well-founded interest in improved versions
or sharpened inequalities of Schur and Bernstein types for various subclasses of polynomials. An important class of interest is the Bernstein polynomials of fixed sign, that is, the
so-called “Lorentz class”:
ᏸ := p ∈ ᏼn : p(x)=0 (x ∈ I) .
(1.6)
Our interest here is the Schur-type inequality for the Lorentz class.
2. Previous results for the Bernstein-Lorentz class
For p ∈ ᏸ, that is, for real polynomials p strictly positive (or strictly negative) on the open
unit interval I := (−1,1), a so-called Lorentz representation is possible, see, for example,
[3, volume II, page 83, Aufgabe 49]. Actually, Lorentz [4] considered polynomials having
the representation
p(x) =
d
ak (1 − x)k (1 + x)d−k
ak ≥ 0 (k = 1,...,d) ,
(2.1)
k =0
where d ∈ N could be any natural number depending on p ∈ ᏼ. Polynomials of this
type were used by Lorentz [4] and others in various questions of approximation theory
such as approximation by incomplete polynomials, shape preserving approximation, and
Szilárd Gy. Révész 3
polynomials with integer coefficients. Regarding these, we refer to [1, 2, 4–8] and the
references therein.
The study of the Lorentz class (1.6), the Lorentz representation (2.1), and the “Lorentz
degree” d = d(p)—defined as the minimal possible degree d of such a representation
of the polynomial—is connected to another basic area of interest. Namely, the general
idea behind the representation (2.1) is to exhibit the nonnegative polynomial p ∈ ᏼn
as the positive (nonnegative) linear combination of positive (nonnegative) polynomials
qkd (x) := (1 − x)k (1 + x)d−k .
The positive elements qkd form a basis of ᏼd , and (2.1) is a positive representation,
that is, a representation with all coefficients ak ≥ 0. Do all p ≥ 0, p ∈ ᏼd have the positive
representation (2.1)? It is easy to see that the answer to this question is negative. However,
such questions lead to other interesting problems, and the whole issue is a vast field of
investigations embedded into the theory of Banach lattices and positive bases, see, for
example, [9–11]. In particular, these general results show that ᏼn does not have a positive
basis at all, and, moreover, any subspace of ᏼn with a positive basis has dimension at most
n/2
. For these questions, we refer the reader to [12].
Another related matter is the theory of positive operators, in particular, Bernstein operators:
Bn ( f ,x) :=
n 2k − n n f
k =0
n
k
(1 − x)k (1 + x)n−k .
(2.2)
Clearly, Bn maps C(I) to ᏼn , and for f ≥ 0, Bn ( f ) ≥ 0, that is, Bn ( f ) ∈ ᏸ+ , where
:= { p ∈ ᏼ : p|I > 0}. The Bernstein operators are used extensively in the theory of
approximation, in particular, for their shape-preserving properties.
Now if p ∈ ᏼn , p ≥ 0 were a fixed-point of Bn , by comparing (2.1) and (2.2), we could
deduce p ∈ ᏸ+ and d(p) ≤ n. Since not all p ∈ ᏼn ∩ ᏸ+ have Lorentz degree d(p) ≤ n,
we see that Bn |ᏼn ∩ᏸ+ cannot be identity. In other words, it turns out that the Bernstein
operator is not a projection on the set ᏼn . This in turn explains the shortcomings with
respect to order of approximation compared to projective operators (like, e.g., the de la
Vallée Poussin operator).
Erdélyi and Szabados proved Schur- and Bernstein-type inequalities for these polynomials using their Lorentz degree instead of the ordinary algebraic degree. That brings into
focus the question of determining, or at least estimating, the Lorentz degree.
However, estimating the Lorentz degree of a polynomial p ∈ ᏸ is usually a complicated
matter. There are estimates of d(p) in terms of the zero-free region of p described in
[5, 13]. Here we restrict our attention to the most appealing result of this type, attributed
to Lorentz, see [13, 14].
ᏸ+
Theorem 2.1 (Lorentz). Let p ∈ ᏸ. If p|D =0, then one has d(p) = deg (p), the ordinary
degree.
The reason to pursue estimates of the Lorentz degree is that there are variants of
Schur’s (and also Bernstein’s and Markov’s) inequalities to Lorentz polynomials with the
Lorentz degree taking over the role of the ordinary algebraic degree. Erdélyi and Szabados
[13] (see also [1, E.14, page 436]) have proved the following theorem.
4
Journal of Inequalities and Applications
Theorem 2.2 (Erdélyi-Szabados). Let p ∈ ᏸ have Lorentz degree d = d(p). Then for any
α > 0 one has
p ≤
(d + 2α)d+2α 2 α
p ∈ ᏼ ∩ ᏸ,d = d(p) .
d+α p(x) 1 − x
α
(4α) (d + α)
(2.3)
α
Observe
√ that here the “Schur constant” is of the order d , and in case α = 1/2, it becomes d, which is a considerable improvement compared to (1.2) provided d is not
much larger than n. In particular, combining Theorems A and B gave to Erdélyi and Szabados [13] the following theorem.
Theorem 2.3 (Erdélyi-Szabados). Let p ∈ ᏸ ∩ ᏼn and assume that p|D =0. Then for any
α > 0 one has
p ≤
(n + 2α)n+2α p(x) 1 − x2 α .
(4α)(n + α)n+α
(2.4)
Erdélyi and Szabados exhibit the sharpness of (2.4) as well. They also note that their
method is bound to use positivity of p ∈ ᏸ and the result of Theorem A for the Lorentz
degree, while formally their end result does not refer to Lorentz degree at all: the formulation of their results on these inequalities does not even need the notion of Lorentz degree
and Lorentz representation for this special subclass. Hence they comment: “A direct proof
of this statement would be interesting.”
3. Results
Here we will show that it is possible to obtain Theorem C directly, using only nonvanishing of p on d. Moreover, we investigate the similar questions for complex polynomials,
where the above convex representation is not available. It turns out that the Schur-type
inequalities extend to the complex case unchanged for all p ∈ ᏼcn (and thus without assuming any positivity property at all), with the only assumption of nonvanishing in D.
This is somewhat unexpected, as an example of Halász already established that as regards
Bernstein- and Markov-type inequalities, only worse estimates can be obtained for complex polynomials [15], [1, page 447].
We formulate the following.
Theorem 3.1. Let φ(t) : [0,1]→(0, ∞) be any decreasing, continuous weight function. Consider a polynomial p ∈ ᏼcn and suppose that p|D =0. Then
p(x) ≤
2n
2n
p(x)φ |x| = n
1 + |x| n φ |x| p(x)φ |x|
(1 + a) φ(a)
(3.1)
with a ∈ [0,1] being any point of maximum of φ(t)(1 + t)n on [0,1]. Moreover, equality
occurs only for the polynomials p(x) = c(1 ± x)n with c ∈ C arbitrary.
Corollary 3.2. Let p ∈ ᏼcn and suppose that p|D =0. Then (2.4) holds true for any parameter α > 0. Moreover, equality occurs only for the polynomials p(x) = c(1 ± x)n with c ∈ C
arbitrary.
Szilárd Gy. Révész 5
4. Proof of Theorem 3.1
Lemma 4.1. For arbitrary z ∈ D and 0 < a < 1 one has
z −x
2
z −a ≤ 1+a
∀x ∈ [a,1] .
(4.1)
Moreover, equality can occur in (4.1) only if z = −1 and x = 1.
Proof. In case Rz ≥ (1 + a)/2, we have |z − x| ≤ max x∈[a,1] |z − x| = max (|z − a|, |z −
1|) = |z − a|, because for Rz ≥ (1 + a)/2 also |z − a| ≥ |z − 1| holds. Hence in this case
(4.1) follows even with 1 < 2/(1 + a) on the right-hand side.
In case Rz < (1 + a)/2, we have, similarly to the above, |z − x| ≤ |z − 1|. Let us consider
C
assuming real
now the map f (z) := (z − 1)/(z − a). This is a rational linear map of C→
values on R, hence it is also symmetric to the real axis. Moreover, f maps the set of
all circles and lines to itself, f (∞) = 1, f (1) = 0, f (a) = ∞, and f (−1) = 2/(1 + a). It
follows that the image of the unit circle C = ∂D will be the circle K symmetric to R and
going through the points 0 and 2/(1 + a), that is, the circle with center 1/(1 + a) and
radius r := 1/(1 + a). Moreover, the domain outside of D is mapped onto the interior
disk B of K = ∂B, since f (∞) = 1 ∈ (0,2/(1 + a)) ⊂ B. However, B ⊂ D(0,2/(1 + a)), the
disk centered at the origin and of radius 2/(1 + a). Thus, for all z ∈ D the image satisfies
f (z) ∈ B and therefore | f (z)| ≤ 2/(1 + a). Consequently, we conclude in this case again
that
z −x z −1 2
≤
z − a z − a = f (z) ≤ 1 + a .
(4.2)
Moreover, in case Rz ≥ (1 + a)/2, there holds a strict inequality, and in case Rz < (1 +
a)/2, |z − x| = |z − 1| entails x = 1, and | f (z)| = 2/(1 + a) entails z = −1. Thus, the asser
tion regarding case of equality follows, too.
Remark 4.2. The above proof follows mapping properties and deduces the estimate from
geometrical facts. We thank the referee for pointing out an even quicker, purely algebraic
proof, which, however, does not reveal how one may find the right assertion. That explains our choice of keeping the first argument, while the reader may recover himself the
latter.
Proof of Theorem 1. Take any parameter 0 < a < 1, and consider the polynomial
Pn (x) := (1 + x)n .
Plainly, for any p(x) =
n
j =1 (x − z j ),
(4.3)
where for all j = 1,...,n we have |z j | ≥ 1, we have
sup p(x)
x∈[a,1]
n
n x − zj 2 n
· p(a) ≤
p(a) = Pn (1) p(a);
x
− z j = sup
= sup a
−
z
1
+
a
P
(a)
j
n
a≤x≤1 j =1
a≤x≤1 j =1
(4.4)
6
Journal of Inequalities and Applications
hence
sup p(x) ≤
x∈[a,1]
Pn (1) p(a)φ(a) ≤ Pn (1) p(x)φ(x)
[0,1] .
Pn (a)φ(a)
Pn (a)φ(a)
(4.5)
On the other hand, for 0 ≤ x ≤ a we trivially have
p(x) ≤
1 φ(x)p(x) ≤ 1 p(x)φ(x)
[0,1] .
φ(a)
φ(a)
(4.6)
Combining (4.5) and (4.6) we obtain
sup p(x) ≤
x∈[0,1]
Pn (1) p(x)φ(x)
[0,1] ,
Pn (a)φ(a)
(4.7)
and applying this also to p(−x) we finally get
p[−1,1] ≤
Pn (1) p(x)φ |x| [−1,1] .
Pn (a)φ(a)
(4.8)
Note that (4.8) actually means also that
max
P n p[−1,1]
c
: p ∈ ᏼn , p|D =0 = ,
p(x)φ |x| Pn (x)φ(x)
[−1,1]
[0,1]
(4.9)
because (4.8) holds for all 0 <a <1 and hence the maximum can be taken all over 0 <a <1.
Suppose now that we have equality in the statement of the theorem, that is, in (3.1).
Since (4.8) was a consequence of (4.7) and its application to p(−x), case of equality occurs
only if (4.7) holds with equality either for p(x) or for p(−x). Suppose, for example, that
we have equality in (4.7) for p(x), which implies equality also in (4.4) and (4.5) as well.
Equality in (4.4) in turn yields | p(x0 )| = (Pn (1)/Pn (a))| p(a)| = (2/(1 + a))n | p(a)| for the
maximum point x0 ∈ [a,1] of p, and now the equality part of the assertion of Lemma 4.1
implies x0 = 1 and z j = −1( j = 1,...,n) for all roots of p. But this proves p(x) = c(1 + x)n ,
and in case of equality for p(−x), we similarly obtain p(x) = c(1 − x)n . This concludes the
proof.
Proof of Corollary 2. Computing the norm on the right-hand side of (4.9) for φ(x) =
α
(1 − x2 ) —that is, equivalently, taking a = n/(n + 2α) in (4.8)—yields
p
2n
(n + 2α)n+2α
α ≤ =
,
α
n
p(x) 1 − x2 (n + α)n+α (4α)α
1 + n/(n + 2α) 1 − n2 /(n + 2α)2
which proves (2.4).
(4.10)
5. Remarks and examples
Comparing our proof with that of Erdélyi and Szabados, we can realize that the standard approach is to make use of the convex combination (2.1). Denote the set of positive
Szilárd Gy. Révész 7
Lorentz polynomials of Lorentz degree not exceeding d, or ordinary degree not exceeding n by ᏸ+d and by ᏸ+n , respectively. It is obvious that ᏸ+d , ᏸ+n , and ᏸ+ are convex sets.
Using convexity of ᏸ+d , that is, working out proofs for the basis functions qk,d (x) and
then adding the results, is a convenient method for real Lorentz polynomials. However,
departing real polynomials, we necessarily need complex coefficients, and for ᏼc ∩ ᏸ
similar arguments do not work. It turns out that not even the set
ᏼcn (D) := p ∈ ᏼcn : p|I > 0, p|D =0
(5.1)
is convex; hence convex combinations cannot be used directly in this setting.
√
Example 5.1. Let 0 < a < 1, w := a + i 1 − a2 , and consider the polynomials
p(x) := (1 − x) x2 − 2ax + 1 = (1 − x)(x − w)(x − w)
= −x3 + (1 + 2a)x2 − (2a + 1)x + 1
q(x) := 1 + x + x2 + x3 = (1 + x) 1 + x2 = (x + 1)(x + i)(x − i).
(5.2)
Then p, q ∈ ᏼc3 (D), but for r := (1/2)p + (1/2)q one has r ∈ ᏼc3 (D), hence ᏼc3 (D) is not
convex.
Indeed, both p and q have zeros of absolute value 1 only, so they belong to ᏼc3 (D).
Moreover, for
r(x) =
p(x) + q(x)
= (1 + a)x2 − ax + 1,
2
(5.3)
we obviously have r ∈ ᏸ+ (ᏸ+ is convex). On the other hand, the roots of r(x) are
√
a ± a2 − 4(1 + a) a ± i 4 + 4a − a2
=
.
x1,2 =
2(1 + a)
2(1 + a)
(5.4)
Observe that 4 + 4a − a2 > 4 > 0 for all a ∈ (0,1). Now we can compute
2
2
x1,2 2 = a + 4 + 4a − a =
2
4(1 + a)
√
1
≤ 1,
1+a
(5.5)
hence |x1,2 | = 1/ 1 + a < 1, x1,2 ∈ D, and r ∈ ᏼcn (D) for any n ∈ N.
Note that in this example both p and q have degree 3, and by Theorem A and p, q ∈
ᏼc3 (D) their Lorentz degree is 3. Consequently, by convexity of ᏸ+ and ᏸ+d , we must
have d(r) ≤ 3, while d(r) ≥ deg r = 2. To decide the exact value of d(r), note that (1 + x)2 ,
1 − x2 , and (1 − x)2 form a basis of ᏼ2 , and easy linear algebra gives r(x) = (1/2)(1 + x)2 +
((1 + a)/2)(1 − x)2 − (a/2)(1 − x2 ), whence the unique degree 2 representation is not positive and the Lorentz degree cannot be 2. Actually, d(r)=deg r already follows from [13,
Theorem 2(ii)] or [13, Proposition, page 117]. Whence d(r) = 3, and the corresponding
representation is easily obtained from those of p and q.
8
Journal of Inequalities and Applications
The following comment was offered by Tamás Erdélyi.
Remark 5.2 (Erdélyi). As regards Schur’s inequality, we have a better than general bound
(2.4) at least for the class ᏼcn (D). In fact, this can also be obtained from the real case, that
is, from Theorems B and A, independently of Theorem 3.1 or Corollary 3.2.
Indeed, let p ∈ ᏼcn such that p(z)=0 for z ∈ D. Consider also
p(z) := p(z) =
n
z − zj
p(z) :=
j =1
n
z − zj
(5.6)
j =1
and take p∗ (z) := p(z) p(z). Obviously, p∗ ∈ ᏼ2n and p∗ ∈ ᏸ+2n , too. Applying Theorem
B with power α∗ := 2α to p∗ of degree n∗ := 2n, we get
p2 = p∗ = p p ≤
(2n + 4α)2n+4α
2 2α 2α
(2n+2α) p(x) p(x) 1 − x
(8α) (2n + 2α)
2
(n + 2α)n+2α 2 α
.
=
α
n+α p(x) 1 − x
(4α) (n + α)
(5.7)
However, for the Bernstein and Markov inequalities in the generality of complex polynomials not vanishing in D, we have substantially worse factors, see [1, page 474] and
[15]. The example of Halász below shows what we can expect at most.
Example 5.3 (Halász). Let m be chosen as [(n − 1)/2], so that 2m + 1 ≤ n ≤ 2m + 2. Define
the deg P = n polynomial P as
P(z) := (z − 1)
m
2
z − e2πi j/(2m+1) .
(5.8)
j =1
Then P |D =0, P D = 2 = | p(−1)|, and p(−1) cnlog n. Moreover, for any x ∈ [−1,1],
we also have p(x) > cnlog (e/(1 − x2 )) whenever this is smaller than cnlog n.
Consequently, no better bound, than c min (nlogn;nlog(e/(1 − x2 )), is valid in the
Markov and Bernstein inequalities, even if we restrict to ᏼcn (D).
The (essentially standard) calculation showing these lower estimates can be found, for
example, in [15] or [1, page 447]. These are indeed the right factors as the corresponding
upper estimation is proved, for example, in [15].
A standard way of proving Markov-type inequalities is to combine Bernstein inequalities with Schur inequalities. Of course, to get a sharp Markov estimate we must combine
sharp Bernstein and sharp Schur inequalities as well. Thanks to the general form (with
any monotone φ(x)) of our formulation of the Schur-type inequality Theorem 3.1, here
we can indeed deduce the Markov bound from the corresponding Bernstein inequality.
Indeed, the known Bernstein-type estimate (see [15, Theorem 2.1]) says that
p(x) ≤ nlog
e
1 − x2
|x| < 1, p ∈ ᏼcn (D) ,
(5.9)
Szilárd Gy. Révész 9
and applying the Schur inequality (3.1) to p(x) and φ(x) := log −1 (e/(1 − x2 )) we obtain
2n
2n
p(x) ≤ (1 + x)n φ(x) φ(x)p(x) ≤ 1 + x0 n φ x0 φ(x)p(x)
(5.10)
with arbitrary x0 ∈ I. Choosing x0 := 1 − 2/n, say, we thus obtain p ≤ C lognφp
and this can be estimated by the above Bernstein inequality (5.9) as ≤ Cnlog n.
Note that given the logarithmic weight in the complex case, restricting to weights
α
(1 − x2 ) would bring by itself the loss of the possibility of this deduction.
Acknowledgments
The author was supported in part by the Hungarian National Foundation for Scientific
Research, Project no. T-049301 and K-61908. This work was accomplished during the author’s stay in Paris under his Marie Curie fellowship, contract no. MEIF-CT-2005-022927.
References
[1] P. Borwein and T. Erdélyi, Polynomials and Polynomial Inequalities, vol. 161 of Graduate Texts in
Mathematics, Springer, New York, NY, USA, 1995.
[2] G. V. Milovanović, D. S. Mitrinović, and Th. M. Rassias, Topics in Polynomials: Extremal Problems, Inequalities, Zeros, World Scientific, River Edge, NJ, USA, 1994.
[3] Gy. Pólya and G. Szegö, Aufgaben und Lehrsätze aus der Analysis, Vol. II, vol. 20 of Die
Grundlehren der matehmatischen Wissenschaften in Einzeldarstellungen, Springer, Berlin, Germany, 1925.
[4] G. G. Lorentz, “The degree of approximation by polynomials with positive coefficients,” Mathematische Annalen, vol. 151, no. 3, pp. 239–251, 1963.
[5] T. Erdélyi, “Estimates for the Lorentz degree of polynomials,” Journal of Approximation Theory,
vol. 67, no. 2, pp. 187–198, 1991.
[6] K. A. Kopotun, “Uniform estimates of monotone and convex approximation of smooth functions,” Journal of Approximation Theory, vol. 80, no. 1, pp. 76–107, 1995.
[7] K. A. Kopotun, D. Leviatan, and I. A. Shevchuk, “Convex polynomial approximation in the
uniform norm: conclusion,” Canadian Journal of Mathematics, vol. 57, no. 6, pp. 1224–1248,
2005.
[8] K. Kopotun, D. Leviatan, and I. A. Shevchuk, “Coconvex approximation in the uniform norm:
the final frontier,” Acta Mathematica Hungarica, vol. 110, no. 1-2, pp. 117–151, 2006.
[9] I. A. Polyrakis, “Finite-dimensional lattice-subspaces of C(Ω) and curves of Rn ,” Transactions of
the American Mathematical Society, vol. 348, no. 7, pp. 2793–2810, 1996.
[10] I. A. Polyrakis, “Lattice-subspaces of C[0,1] and positive bases,” Journal of Mathematical Analysis
and Applications, vol. 184, no. 1, pp. 1–18, 1994.
[11] I. A. Polyrakis, “Minimal lattice-subspaces,” Transactions of the American Mathematical Society,
vol. 351, no. 10, pp. 4183–4203, 1999.
[12] B. Farkas and S. Gy. Révész, “Positive bases in spaces of Polynomials,” preprint, 2007
http://arxiv.org/abs/0708.2883.
[13] T. Erdélyi and J. Szabados, “On polynomials with positive coefficients,” Journal of Approximation
Theory, vol. 54, no. 1, pp. 107–122, 1988.
10
Journal of Inequalities and Applications
[14] J. T. Scheick, “Inequalities for derivatives of polynomials of special type,” Journal of Approximation Theory, vol. 6, no. 4, pp. 354–358, 1972.
[15] T. Erdélyi, “Markov-Bernstein type inequalities for constrained polynomials with real versus
complex coefficients,” Journal d’Analyse Mathématique, vol. 74, pp. 165–181, 1998.
Szilárd Gy. Révész: Alfréd Rényi Institute of Mathematics, Hungarian Academy of Sciences,
P.O. Box 127, 1364 Budapest, Hungary
Email address: revesz@renyi.hu
Download